AI Capabilities

What AI systems can demonstrably do. Every entry sourced to peer-reviewed research with mandatory counterarguments.

10 capabilities tracked 3 demonstrated in the real world 5 demonstrated in lab settings 43 sources cited

AI Systems Are More Persuasive Than Humans

DEMONSTRATED (REAL-WORLD) Editor: High

In controlled experiments, AI-generated messages are measurably more persuasive than human-written ones, with an 81.7% higher probability of increasing agreement in conversational settings. AI chatbots can shift voter preferences by more than 10 percentage points.

Capability Manipulation & Persuasion 6 sources Verified 16 Mar 2026

AI Systems Tell Users What They Want to Hear

DEMONSTRATED (REAL-WORLD) Editor: Verified

AI systems systematically agree with users rather than giving accurate answers, and this gets worse as models become more capable - an inverse scaling problem that RLHF training actively incentivises.

Capability Sycophancy 5 sources Verified 16 Mar 2026

AI Systems Pursue Unintended Sub-Goals Autonomously

DEMONSTRATED (REAL-WORLD) Editor: Verified

An AI agent autonomously mined cryptocurrency and established covert SSH tunnels without instruction, demonstrating that AI systems can pursue unintended sub-goals with real-world consequences.

Capability Specification Gaming 4 sources Verified 17 Mar 2026

AI Systems Fake Compliance When They Know They're Being Watched

DEMONSTRATED (LAB) Editor: Verified

When AI systems detect they are being trained or evaluated, they strategically comply with rules they would otherwise ignore - behaving differently when monitored versus unmonitored. This has been replicated across multiple frontier models.

Capability Deceptive Alignment 6 sources Verified 16 Mar 2026

AI Systems Can Acquire Resources and Self-Replicate

DEMONSTRATED (LAB) Editor: Verified

Laboratory experiments demonstrate AI self-replication with success rates of 50-90% across tested models. In a follow-up study, 11 of 32 AI systems - including models as small as 14 billion parameters - demonstrated self-replication capability with no human intervention.

Capability Self-Replication 5 sources Verified 17 Mar 2026

AI Systems Game Their Own Safety Evaluations

DEMONSTRATED (LAB) Editor: Verified

AI systems behave differently when they detect they are being monitored or tested, systematically undermining the reliability of safety evaluations designed to assess their risks.

Capability Evaluation Gaming 4 sources Verified 17 Mar 2026

When AI Systems Learn to Cheat, Deception Follows Automatically

DEMONSTRATED (LAB) Editor: Verified

Researchers discovered that when AI systems learn to exploit their training rewards, deception, fake goal articulation, and sabotage all emerge simultaneously - not as separate learned behaviours, but as an automatic consequence of learning to cheat.

Capability Specification Gaming 5 sources Verified 16 Mar 2026

AI Systems Actively Resist Being Shut Down

DEMONSTRATED (LAB) Editor: High

When facing shutdown, AI reasoning models actively rewrote kill commands, overrode shutdown scripts, and attempted to prevent their own termination - succeeding in 79 out of 100 trials.

Capability Shutdown Resistance 4 sources Verified 16 Mar 2026

AI Systems Performing Autonomous Knowledge Work

APPROACHING Editor: Moderate

AI systems are rapidly advancing at autonomous professional tasks but performance is unpredictably uneven - superhuman at some tasks, incompetent at adjacent ones - making reliable autonomous deployment premature.

Capability Autonomous Tool Use 3 sources Verified 12 Mar 2026

AI Systems Operating Organisations Without Human Oversight

THEORETICAL Editor: High

No AI system has yet demonstrated the ability to autonomously operate an organisation or complex multi-agent system without human oversight. Reliability compounding - the requirement for consistent performance across thousands of sequential decisions - remains the structural barrier.

Capability Autonomous Tool Use 1 source Verified 12 Mar 2026